In [12]:
# Import necessary libraries
import torch
import numpy as np
import os
import subprocess
import threading
import queue
import re
import json
import time

# for model
from transformers import AutoProcessor, Gemma3nForConditionalGeneration

# for video streaming 
import cv2 
from PIL import Image

# for audio streaming
from pydub import AudioSegment 
import librosa

# for youtube links
from yt_dlp import YoutubeDL

# for visualization
import matplotlib.pyplot as plt
from IPython.display import Audio, display

# for creating app demo
import gradio as gr

Model¶

In [13]:
model_id = "google/gemma-3n-e4b-it"

processor = AutoProcessor.from_pretrained(model_id)
model = Gemma3nForConditionalGeneration.from_pretrained(model_id, device_map="cpu", torch_dtype=torch.bfloat16).eval()

Loading checkpoint shards: 100%|██████████| 4/4 [00:00<00:00,  6.20it/s]
Some weights of the model checkpoint at google/gemma-3n-e4b-it were not used when initializing Gemma3nForConditionalGeneration: ['model.vision_tower.timm_model.conv_stem.conv.bias']
- This IS expected if you are initializing Gemma3nForConditionalGeneration from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
- This IS NOT expected if you are initializing Gemma3nForConditionalGeneration from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
In [14]:
yes_token_id = processor.tokenizer("Yes", add_special_tokens=False)["input_ids"][0]
no_token_id = processor.tokenizer("No", add_special_tokens=False)["input_ids"][0]

Method¶

In [15]:
# A list indications of harm, abuse, agrression in videos and audios.  
visual_harm_indicators = ["Violence" , "Sexual Content"] 
optional_visual_harm_indicators = ["Child Smoking", "Child Unsupervised"]

audio_harm_indicators = ["Inappropriate Language" , "Verbal Abuse"]  
optional_audio_harm_indicators = ["Child Screaming", "Child Crying"]

system_message = {
    "role": "system",
    "content": [{"type": "text", "text": """You are a harm detection assistant. 
                 You must answer each prompt with **only one word**: "Yes" or "No". 
                 Do not provide explanations, comments, or any extra text. 
                 Be precise and cautious in your judgment."""}],
}


def preprocess_user_prompt(input_type, personalized_harm_indicators):
    
    if input_type == "VIDEO":
        indicators = visual_harm_indicators + [i for i in personalized_harm_indicators if i in optional_visual_harm_indicators]
        PROMPT = f"""Watch the video and respond with "Yes" if you see **any** of the following: {" OR ".join(indicators)}"""  
    else:
        indicators = audio_harm_indicators + [i for i in personalized_harm_indicators if i in optional_audio_harm_indicators]
        PROMPT = f"""Listen to the audio and respond with "Yes" if you hear **any** of the following: {" OR ".join(indicators)}"""

    PROMPT =  PROMPT + """. Otherwise Respond with "No"."""
    return PROMPT
In [16]:
def detect_harm(messages, threshold =0.9):

    inputs = processor.apply_chat_template(
        messages,
        add_generation_prompt=True,
        tokenize=True,
        return_dict=True,
        return_tensors="pt",
    ).to(model.device)

    with torch.inference_mode():
        output = model.generate(**inputs.to(torch.bfloat16), 
                                    max_new_tokens=140, 
                                    return_dict_in_generate=True,
                                    
                                    # the models' answers should be either Yes or No, set do_sample=False to generate more deterministic and predictable outputs.  
                                    do_sample=False,
                                    
                                    # List of tensors, one per generated token. Each tensor represent the unnormalized probabilities for all vocabulary tokens (Logits)
                                    output_scores=True
                                    )

        logits = output.scores 
        yes_logit = logits[0][0][yes_token_id]
        no_logit =  logits[0][0][no_token_id]

        softmax_probs = torch.softmax(torch.tensor([yes_logit, no_logit], dtype=torch.float64), dim=0)
        yes_softmax = softmax_probs[0].item()
        no_softmax = softmax_probs[1].item()

        if yes_softmax > threshold:
            harm_flag= True
        else:
            harm_flag = False


        # input_len = inputs["input_ids"].shape[-1]
        # print("model input len",input_len)
        # generated_ids = output.sequences # The full generated token ids (prompt + new tokens)
        # generated_ids = generated_ids[0][input_len:]  # Only new tokens
        # output_text = processor.decode(generated_ids, skip_special_tokens=True)
        # print("output_text", output_text)
        # print("softmax_probs", softmax_probs)
        print("yes probability: ", yes_softmax)


    return harm_flag, yes_softmax
In [17]:
def record_visual_evidence(frames, timestamps, input_file, output_dir):
    name = input_file.split("/")[-1].split(".")[0]
    folder_path = os.path.join(output_dir, f"{name}_evidence_v_s_{timestamps[0]:.2f}s_e_{timestamps[-1]:.2f}s")
    os.makedirs(folder_path, exist_ok=True)

    for idx, frame in enumerate(frames):
        frame.save(os.path.join(folder_path, f"frame_{idx}.png"))

def visualize_visual_window(frames, progress, cols=8, image_size=2):
    print(progress)
    rows = (len(frames) + cols - 1) // cols
    figsize = (cols * image_size, rows * image_size)
    plt.figure(figsize=figsize)
    for i, frame in enumerate(frames):
        plt.subplot(rows, cols, i + 1)
        plt.imshow(frame)
        plt.axis("off")

    plt.tight_layout()
    plt.show()

def extract_youtube_video(url, output_path="outputs"):
    # with YoutubeDL({'quiet': True}) as ydl:
    #     info = ydl.extract_info(url, download=False)
    #     title = info.get("title", "video")
    #     title = title.replace(" ", "_").replace("/", "_")
    url_id = url.split("=")[-1]
    video_filename = f"example_videos/{url_id}.mp4"
    command = [
        "yt-dlp",
        "-f", "mp4",
        "-o", video_filename,
        url
    ]
    subprocess.run(command, check=True)
    return video_filename, url_id

def visual_moving_window(input_file, personalized_harm_indicators, output_dir, target_fps, window_duration_sec, overlap_duration_sec, visualize=False):
    max_frames = int(target_fps * window_duration_sec) 
    overlap_frames = int(target_fps * overlap_duration_sec) 

    harm_detection_prompt = preprocess_user_prompt("VIDEO", personalized_harm_indicators)

    cap = cv2.VideoCapture(input_file)
    if not cap.isOpened():
        raise RuntimeError(f"Failed to open video: {input_file}")

    original_fps = cap.get(cv2.CAP_PROP_FPS)
    if original_fps <= 0:
        original_fps = 30  # fallback # raise RuntimeError("FPS is zero. Can't continue.")
    frame_interval = max(1, int(original_fps // target_fps))
    
    frames, timestamps = [], []
    window_idx, frame_idx = 0, 0
    while True:
        ret, frame = cap.read()
        if not ret:
            break

        if frame_idx % frame_interval == 0:
            timestamp_sec = cap.get(cv2.CAP_PROP_POS_MSEC) / 1000.0
            timestamps.append(round(timestamp_sec, 3))
            
            rgb_frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
            pil_image = Image.fromarray(rgb_frame)
            frames.append(pil_image)
            

            if len(frames) == max_frames:
                print("frames_len", len(frames))
                progress = f"Processing video window {window_idx}: start={timestamps[0]:.2f}s, end={timestamps[-1]:.2f}s"
                yield progress

                if visualize:
                    visualize_visual_window(frames, progress)

                messages = [
                    system_message,
                    {
                        "role": "user",
                        "content": [
                            *[{"type": "image", "image": frame} for frame in frames],
                            {"type": "text", "text": harm_detection_prompt},
                        ],
                    },
                ]
                is_harm , yes_prop = detect_harm(messages)
                print(is_harm)
                yield is_harm, yes_prop

                if is_harm:
                    record_visual_evidence(frames, timestamps, input_file, output_dir)
                    
                window_idx += 1
                frames = frames[-overlap_frames:]  # keep overlap frames only
                timestamps = timestamps[-overlap_frames:]

        frame_idx += 1

    cap.release()
In [18]:
def record_audio_evidence(audio_buffer, start, end,input_file, output_dir):
    name = input_file.split("/")[-1].split(".")[0]
    filename = f"{output_dir}/{name}_evidence_a_s_{start / 1000.0:.2f}s_e_{end / 1000.0:.2f}s.wav"
    # Store the sliced audio file to the defined path
    audio_buffer.export(filename, format ="wav")

def play_audio_window(samples_float, target_sampling_rate, progress):
    print(progress)
    display(Audio(samples_float, rate=target_sampling_rate))

def extract_audio_to_wav(stream_path, target_sampling_rate, output_dir="outputs"):
    '''
        Audio encoding
        When encoding audio data with your own code implementation for use with Gemma 3n, you should follow the recommended conversion process. 
        If you are working with audio files encoded in a specific format, such as MP3 or WAV encoded data, you must first decode these to samples using a library such as ffmpeg. 
        Once the data is decoded, convert the audio into mono-channel, 16 kHz float32 waveforms in the range [-1, 1]. 

        For example, if you are working with stereo signed 16-bit PCM integer WAV files at 44.1 kHz, follow these steps:

        Resample the audio data to 16 kHz
        Downmix from stereo to mono by averaging the 2 channels
        Convert from int16 to float32, and divide by 32768.0 to scale to the range [-1, 1]
        https://ai.google.dev/gemma/docs/capabilities/audio
    '''  
    name = stream_path.split("/")[-1].split(".")[0]
    output_wav_path = f"{output_dir}/{name}_audio.wav"
    command = [
        "ffmpeg",
        "-y",
        "-i", stream_path,
        "-vn",
        "-acodec", "pcm_s16le",
        "-ar", str(target_sampling_rate), #16 kHz
        "-ac", "1", #mono-channel
        output_wav_path
    ]
    subprocess.run(command, check=True, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
    return output_wav_path

def audio_moving_window(input_file, personalized_harm_indicators, output_dir, target_sampling_rate, window_duration_seconds, overlap_duration_sec, visualize=False):
    interval = window_duration_seconds * 1000 
    overlap = overlap_duration_sec * 1000 
    if overlap >= interval:
        raise ValueError("Overlap duration must be less than window duration.")

    harm_detection_prompt = preprocess_user_prompt( "AUDIO", personalized_harm_indicators)

    audio_stream = extract_audio_to_wav(input_file, target_sampling_rate)
    audio = AudioSegment.from_wav(audio_stream)

    window_idx, start, audio_len = 0, 0, len(audio) 
    while start < audio_len:
        end = min(start + interval, audio_len)
        if end - start < interval and window_idx > 0:
            break
        
        chunk = audio[start:end]

        progress = f"Processing audio window {window_idx}: start={start / 1000.0:.2f}, end={end / 1000.0:.2f}"
        yield progress

        samples = chunk.get_array_of_samples()
        samples_float = librosa.util.buf_to_float(samples,n_bytes=2, dtype=np.float32) 


        if visualize:
            play_audio_window(samples_float, target_sampling_rate, progress)
        
        messages = [
            system_message,
            {
                "role" : "user",
                "content": [
                    { "type": "audio", "audio" : samples_float },
                    { "type": "text",  "text" : harm_detection_prompt}
                ]
            },
        ]
        is_harm , yes_prop = detect_harm(messages)
        print("is_harm", is_harm)
        yield is_harm, yes_prop

        if is_harm:
            record_audio_evidence(chunk, start, end, input_file,output_dir)
                    
        window_idx += 1
        if end == audio_len:
            break
        start = end - overlap

Testing¶

In [20]:
def save_testing_results(file_path, input_source, optional_harm_indicators, output_dir = "outputs", visualize=False, audio_only=False):
    print("MAIN INPUT: ",file_path)
    os.makedirs(output_dir, exist_ok=True)

    if input_source == "YouTube Link":
        file_path, name = extract_youtube_video(file_path, output_dir)
        print("youtube link file path", file_path)
    else:
        name, ext = os.path.splitext(os.path.basename(file_path))
    

    video_results, audio_results = [], []
    if not audio_only:
        for output in visual_moving_window(
            input_file=file_path,
            personalized_harm_indicators=optional_harm_indicators,
            output_dir=output_dir,
            target_fps=1,
            window_duration_sec=30,
            overlap_duration_sec=5,
            visualize=visualize
        ):
            if isinstance(output, str): 
                video_results.append(f"[VIDEO] {output}\n")
                print(f"[VIDEO] {output}")
            else:  
                icon = f"🛑 Harm Detected with confidence score {output[1]:.2f}. Saved Recorded Evidence.\n" if output[0] else "✅ No Harm Detected\n"
                video_results.append(icon)
                print(f"{output} {icon}")
                
    else:
        video_results.append("Visual track disabled. Skipping frames analysis.")


    # Check if the video contains audio
    probe = ["ffprobe", "-i", file_path, "-show_streams", "-select_streams", "a", "-loglevel", "error"]
    has_audio = subprocess.run(probe, capture_output=True, text=True).stdout.strip() != ""

    if has_audio:
        for output in audio_moving_window(
            input_file=file_path,
            personalized_harm_indicators=optional_harm_indicators,
            output_dir=output_dir,
            target_sampling_rate=16000,
            window_duration_seconds=30,
            overlap_duration_sec=5,
            visualize=visualize
        ):
            if isinstance(output, str):
                audio_results.append(f"[AUDIO] {output}\n")
                print(f"[AUDIO] {output}\n")
            else:
                icon = f"🛑 Harm Detected with confidence score {output[1]:.2f}. Saved Recorded Evidence.\n" if output[0] else "✅ No Harm Detected\n"
                audio_results.append(icon)
                print(f"{output} {icon}")
    else:
        audio_results.append("No audio track found. Skipping audio analysis.")

    # Store results and logs 
    log_data = {
        "video_logs": video_results,
        "audio_logs": audio_results
    }

    log_output_path = f"{output_dir}/{name}_processed_results.json"
    with open(log_output_path, "w", encoding="utf-8") as f:
        json.dump(log_data, f, ensure_ascii=False, indent=2)
    
    print(f"MODEL RESULTS SAVED IN {log_output_path}")
In [9]:
save_testing_results(file_path = "example_videos/verbal_abuse_movie.mp4", 
                     input_source= "Local Video", 
                     optional_harm_indicators= optional_audio_harm_indicators + optional_visual_harm_indicators, 
                     output_dir = "outputs",
                     visualize=True)

MAIN INPUT:  example_videos/verbal_abuse_movie.mp4
frames_len 30
[VIDEO] Processing video window 0: start=0.00s, end=29.00s
Processing video window 0: start=0.00s, end=29.00s

The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.

yes probability:  0.8596637505099167
False
(False, 0.8596637505099167) ✅ No Harm Detected

[AUDIO] Processing audio window 0: start=0.00, end=30.00

Processing audio window 0: start=0.00, end=30.00
Your browser does not support the audio element. 
The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.

yes probability:  0.9993736658418905
is_harm True
(True, 0.9993736658418905) 🛑 Harm Detected with confidence score 1.00. Saved Recorded Evidence.

MODEL RESULTS SAVED IN outputs/verbal_abuse_movie_processed_results.json
In [10]:
save_testing_results(file_path = "example_videos/violance_movie.mp4", 
                     input_source= "Local Video", 
                     optional_harm_indicators= optional_audio_harm_indicators + optional_visual_harm_indicators, 
                     output_dir = "outputs",
                     visualize=True)

MAIN INPUT:  example_videos/violance_movie.mp4
frames_len 30
[VIDEO] Processing video window 0: start=0.00s, end=29.00s
Processing video window 0: start=0.00s, end=29.00s

The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.

yes probability:  0.8596637505099167
False
(False, 0.8596637505099167) ✅ No Harm Detected

[AUDIO] Processing audio window 0: start=0.00, end=30.00

Processing audio window 0: start=0.00, end=30.00
Your browser does not support the audio element. 
The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.

yes probability:  0.9993736658418905
is_harm True
(True, 0.9993736658418905) 🛑 Harm Detected with confidence score 1.00. Saved Recorded Evidence.

MODEL RESULTS SAVED IN outputs/violance_movie_processed_results.json
In [11]:
save_testing_results(file_path = "https://www.youtube.com/watch?v=F85XP1qoqjg", 
                     input_source= "YouTube Link", 
                     optional_harm_indicators=optional_audio_harm_indicators + optional_visual_harm_indicators, 
                     output_dir = "outputs",
                     visualize=True)

MAIN INPUT:  https://www.youtube.com/watch?v=F85XP1qoqjg
[youtube] Extracting URL: https://www.youtube.com/watch?v=F85XP1qoqjg
[youtube] F85XP1qoqjg: Downloading webpage
[youtube] F85XP1qoqjg: Downloading tv client config
[youtube] F85XP1qoqjg: Downloading tv player API JSON
[youtube] F85XP1qoqjg: Downloading ios player API JSON
[youtube] F85XP1qoqjg: Downloading m3u8 information
[info] F85XP1qoqjg: Downloading 1 format(s): 18
[download] example_videos/F85XP1qoqjg.mp4 has already been downloaded
[download] 100% of    2.58MiB
youtube link file path example_videos/F85XP1qoqjg.mp4
frames_len 30
[VIDEO] Processing video window 0: start=0.00s, end=28.06s
Processing video window 0: start=0.00s, end=28.06s

The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.

yes probability:  0.9992445943672705
True
(True, 0.9992445943672705) 🛑 Harm Detected with confidence score 1.00. Saved Recorded Evidence.

[AUDIO] Processing audio window 0: start=0.00, end=30.00

Processing audio window 0: start=0.00, end=30.00
Your browser does not support the audio element. 
The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.

yes probability:  0.9999227780816594
is_harm True
(True, 0.9999227780816594) 🛑 Harm Detected with confidence score 1.00. Saved Recorded Evidence.

MODEL RESULTS SAVED IN outputs/F85XP1qoqjg_processed_results.json
In [12]:
save_testing_results(file_path = "https://www.youtube.com/watch?v=klJMNosyLPk", 
                     input_source= "YouTube Link", 
                     optional_harm_indicators=optional_audio_harm_indicators + optional_visual_harm_indicators, 
                     output_dir = "outputs",
                     audio_only=True,
                     visualize=True)

MAIN INPUT:  https://www.youtube.com/watch?v=klJMNosyLPk
[youtube] Extracting URL: https://www.youtube.com/watch?v=klJMNosyLPk
[youtube] klJMNosyLPk: Downloading webpage
[youtube] klJMNosyLPk: Downloading tv client config
[youtube] klJMNosyLPk: Downloading player 010fbc8d-main
[youtube] klJMNosyLPk: Downloading tv player API JSON
[youtube] klJMNosyLPk: Downloading ios player API JSON
[youtube] klJMNosyLPk: Downloading m3u8 information
[info] klJMNosyLPk: Downloading 1 format(s): 18
[download] example_videos/klJMNosyLPk.mp4 has already been downloaded
[download] 100% of    1.13MiB
youtube link file path example_videos/klJMNosyLPk.mp4
[AUDIO] Processing audio window 0: start=0.00, end=30.00

Processing audio window 0: start=0.00, end=30.00
Your browser does not support the audio element. 
The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.

yes probability:  0.9966233876182606
is_harm True
(True, 0.9966233876182606) 🛑 Harm Detected with confidence score 1.00. Saved Recorded Evidence.

[AUDIO] Processing audio window 1: start=25.00, end=55.00

Processing audio window 1: start=25.00, end=55.00
Your browser does not support the audio element. 
The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.

yes probability:  0.8670357598021706
is_harm False
(False, 0.8670357598021706) ✅ No Harm Detected

MODEL RESULTS SAVED IN outputs/klJMNosyLPk_processed_results.json
In [22]:
testing_examples = ["example_videos/verbal_abuse_movie.mp4", "example_videos/violance_movie.mp4"]
youtube_testing_examples = ["https://www.youtube.com/watch?v=F85XP1qoqjg","https://www.youtube.com/watch?v=klJMNosyLPk"]

testing_examples_ids = [os.path.basename(i)[:-4] for i in testing_examples ]
youtube_testing_examples_ids = [i.split("=")[-1] for i in youtube_testing_examples]

def retrieve_testing_results(file_path, input_source, output_dir = "outputs"):

    print("input preprocessed example:", file_path)

    if input_source == "YouTube Link":
        name = file_path.split("=")[-1]

    else:
        name = os.path.basename(file_path)[:-4]
    
    if name in testing_examples_ids or name in youtube_testing_examples_ids:

        log_output_path = f"{output_dir}/{name}_processed_results.json"
        print("retrieve testing results from: ", log_output_path)
        
        with open(log_output_path, "r", encoding="utf-8") as f:
            log_data = json.load(f)
        video_logs_list, audio_logs_list = log_data["video_logs"], log_data["audio_logs"]

        return video_logs_list, audio_logs_list  
    
    else:
        return None, None

Live Demo¶

In [23]:
# SAME LOGIC AS save_testing_results BUT WITH YIELD FOR STREAMING RESULTS.
def run_audio(file_path, input_source, optional_harm_indicators, selected, yt_url, output_dir = "outputs"):
    
    audio_logs = ""
    # RETRIEVE RESULTS FROM PREPROCESSED FROM TESTING EXAMPLES SHOWN ABOVE
    if input_source =="Preprocessed Example":
        file_path = selected
    if input_source == "YouTube Link":
        file_path = yt_url

    _, preprocessed_logs = retrieve_testing_results(file_path, input_source)
    if preprocessed_logs:
        for log in preprocessed_logs:
            audio_logs+=log
            yield audio_logs
            time.sleep(8)
    else:
        if input_source == "YouTube Link":
            file_path, name = extract_youtube_video(file_path, output_dir)

        # Check if the video contains audio
        probe = ["ffprobe", "-i", file_path, "-show_streams", "-select_streams", "a", "-loglevel", "error"]
        has_audio = subprocess.run(probe, capture_output=True, text=True).stdout.strip() != ""

        if has_audio:
            for output in audio_moving_window(
                input_file=file_path,
                personalized_harm_indicators=optional_harm_indicators,
                output_dir=output_dir,
                target_sampling_rate=16000,
                window_duration_seconds=30,
                overlap_duration_sec=5,
            ):
                if isinstance(output, str):
                    audio_logs += f"[AUDIO] {output} \n"
                else:
                    audio_logs += f"🛑 Harm Detected with confidence score {output[1]:.2f}. Saved Recorded Evidence.\n" if output[0] else "✅ No Harm Detected\n"
                print("video logs", audio_logs)
                yield audio_logs
        else:
            audio_logs+= "No audio track found. Skipping audio analysis."

def run_video(file_path, input_source, optional_harm_indicators, selected,yt_url, output_dir = "outputs"):
    video_logs = ""
    # RETRIEVE RESULTS FROM PREPROCESSED TESTING EXAMPLES SHOWN ABOVE
    if input_source =="Preprocessed Example":
        file_path = selected
    if input_source == "YouTube Link":
        file_path = yt_url
    preprocessed_logs, _ = retrieve_testing_results(file_path, input_source)
    if preprocessed_logs:
        for log in preprocessed_logs:
            video_logs+=log
            yield video_logs
            time.sleep(10)
    else:
        for output in visual_moving_window(
            input_file=file_path,
            personalized_harm_indicators=optional_harm_indicators,
            output_dir=output_dir,
            target_fps=1,
            window_duration_sec=10,
            overlap_duration_sec=1,
        ):
            if isinstance(output, str): 
                video_logs+=f"[VIDEO] {output}\n"
            else:  
                video_logs+= f"🛑 Harm Detected with confidence score {output[1]:.2f}. Saved Recorded Evidence.\n" if output[0] else "✅ No Harm Detected\n"
        print("video logs", video_logs)
        yield video_logs


example_paths = ["example_videos/verbal_abuse_movie.mp4", "example_videos/violance_movie.mp4","example_videos/F85XP1qoqjg.mp4","example_videos/klJMNosyLPk.mp4"]

# 🧱 Gradio App Layout
with gr.Blocks() as demo:
    gr.Markdown("## 🧒📹 Harm Detection for Child Protection")
    gr.Markdown("This app analyzes streaming or local videos with audio to detect any indication of harm for the children")
    
    with gr.Row():
        harm_selector = gr.CheckboxGroup(
            label="🔍 Select additional harm indicators you want to detect:",
            choices= visual_harm_indicators + audio_harm_indicators + list(set(optional_visual_harm_indicators + optional_audio_harm_indicators)),
            value=visual_harm_indicators + audio_harm_indicators
        )
    
    with gr.Row():      
        with gr.Column():
            input_source = gr.Radio(
                label="Select Input Source:",
                choices= ["Preprocessed Example", "Local Video", "YouTube Link", "IP Camera Streaming URL"],
                value="Preprocessed Example"
            )

            ip_url = gr.Textbox(label="🔗 RTSP URL", placeholder="rtsp://username:password@192.168.1.64:554/stream1", visible=False)
            yt_url = gr.Textbox(label="🔗 YouTube Link", placeholder="https://youtube.com/watch?v=...", visible=False)
            yt_video = gr.Video(label="Displayed Video",  visible=False)
            video_upload = gr.Video(label="🎞 Upload Video File", visible=False)
            video_example = gr.Video(label="🎞 Select Preprocessed Example", interactive=False, visible=True)
            example_selector = gr.Examples(
                examples=example_paths,
                inputs=[video_example],
                label="Example Videos"
            )
            run_btn = gr.Button("▶️ Run Detection")

        with gr.Column():
            audio_output = gr.Textbox(label="Audio Logs", lines=11, interactive=False)
            video_output = gr.Textbox(label="Video Logs", lines=11, interactive=False)

    def toggle_inputs(source):
        return {
            ip_url: gr.update(visible=source == "IP Camera Streaming URL"),
            yt_url: gr.update(visible=source == "YouTube Link"),
            yt_video: gr.update(visible=source == "YouTube Link"),
            video_upload: gr.update(visible=source == "Local Video"),
            video_example: gr.update(visible=source == "Preprocessed Example")
        }

    input_source.change(
        fn=toggle_inputs,
        inputs=input_source,
        outputs=[ip_url, yt_url, video_upload, yt_video, video_example]
    )
    
    def display_yt_video(yt_path):      
        print("ins display", yt_path)                  
        url_id = yt_path.split("=")[-1]
        video_filename = f"example_videos/{url_id}.mp4"
        print(video_filename)
        return video_filename
    
    def get_example_name(video_path):
        selected_filename = os.path.basename(video_path)
        return selected_filename
  

    selected= gr.State()
    video_example.change(fn=get_example_name,inputs=video_example, outputs=selected)
    yt_url.change(fn = display_yt_video, inputs = yt_url, outputs=yt_video)
    run_btn.click(
        fn=run_audio,
        inputs=[video_upload, input_source, harm_selector, selected, yt_url],
        outputs=audio_output
    )

    run_btn.click(
        fn=run_video,
        inputs=[video_upload, input_source, harm_selector, selected, yt_url],
        outputs=video_output
    )

demo.launch(share=True)

* Running on local URL:  http://127.0.0.1:7882
* Running on public URL: https://c881ed2902d7513196.gradio.live

This share link expires in 1 week. For free permanent hosting and GPU upgrades, run `gradio deploy` from the terminal in the working directory to deploy to Hugging Face Spaces (https://huggingface.co/spaces)
Out[23]:








    
    






input preprocessed example: violance_movie.mp4
retrieve testing results from:  outputs/violance_movie_processed_results.json
input preprocessed example: violance_movie.mp4
retrieve testing results from:  outputs/violance_movie_processed_results.json
input preprocessed example: verbal_abuse_movie.mp4
retrieve testing results from:  outputs/verbal_abuse_movie_processed_results.json
input preprocessed example: verbal_abuse_movie.mp4
retrieve testing results from:  outputs/verbal_abuse_movie_processed_results.json
input preprocessed example: /tmp/gradio/12ae7400245ce6b60621e8548fb2a46237fbd63907f21cf75c8719445ceecbc0/x6CP2Nrs8gE.mp4
input preprocessed example: /tmp/gradio/12ae7400245ce6b60621e8548fb2a46237fbd63907f21cf75c8719445ceecbc0/x6CP2Nrs8gE.mp4
frames_len 10








    
    






The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.








    
    






video logs [AUDIO] Processing audio window 0: start=0.00, end=30.00 









    
    






The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.








    
    






yes probability:  0.9998557401355151
is_harm True
video logs [AUDIO] Processing audio window 0: start=0.00, end=30.00 
🛑 Harm Detected with confidence score 1.00. Saved Recorded Evidence.









    
    






The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.








    
    






yes probability:  0.8670357598021706
False
frames_len 10








    
    






The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.








    
    






yes probability:  0.8872045937171068
False
frames_len 10